System for balance distribution of requests across multiple servers using dynamic metrics

Abstract

A system for distributing incoming client requests across multiple servers in a networked client-server computer environment processes all requests as a set that occur within a given time interval and collects information on the attributes of the requests and the resource capability of the servers to dynamically allocate requests in a set to the appropriate servers upon completion of the time interval. Preferably, a request table collects at least two requests incoming within a predetermined time interval, a request examiner routine analyzes each collected request with respect to at least one attribute, a system status monitor collects resource capability information of each server in a resource table and an optimization and allocation process distributes collected requests in the request table across the multiple servers upon completion of said time interval based on an optimization of potential pairings of the requests in the request table with servers in the resource table.

Description

RELATED APPLICATIONS

This application claims priority to U.S. patent application Ser. No. 09/765,766, filed Jan. 18, 2001, which claims the benefit of U.S. Provisional Application No. 60/176,665, filed Jan. 18, 2000, both of which are incorporated herewith by reference. This application is also related to two co-pending applications that are assigned to the common assignee of the present application, the first of which is entitled “Scalable Internet Engine,” Ser. No. 09/709,820, filed Nov. 10, 2000 and the second of which is entitled “Method and System For Providing Dynamic Host Service Management Across Disparate Accounts/Sites,” Ser. No. 09/710,095, filed Nov. 10, 2000.

FIELD OF THE INVENTION

This invention relates to computers and digital processing systems requiring coordination of multiple digital processing units. In particular, this invention relates to load balancing or distribution of client requests across multiple servers in a networked computing environment.

BACKGROUND OF THE INVENTION

The Internet has become an increasingly useful tool and means of communication to many people. As the popularity of the Internet has increased, traffic to many Internet service provider (ISP) and application service provider (ASP) sites has become so congested at times that many companies have to impose a limit on the number of users using their sites during peak hours. As a result, a significant loss of business for e-business merchants, user dissatisfaction, and a permanent loss of many potential customers occur. According to at least one source, during the 1999 holiday shopping season, 25 percent of all potential online buyers never completed their online purchases because the e-tail sites of interest had either crashed or were simply too slow. The principle cause of these problems in the case of larger sites was and is an inappropriate distribution of the requests of customers or users (clients) among the sites' resources (servers), namely the multiple content and application servers that are responsible for responding to these requests.

Allocating content and application server resources to respond to a large number of client requests can become rather complex in certain circumstances involving multiple servers at a given site. If it is assumed that there is always at least one server available for each new task that arises, resource assignments may be made in an arbitrary manner, making the resource allocation procedure trivial. To satisfy the assumption underlying this approach to resource allocation, it is generally desirable to create a system design that has abundant resources and strives to conserve them to maintain availability and efficient throughput. In this approach, each client request received at a site is handled as an independent event. U.S. Pat. Nos. 6,173,322, 6,070,191, 5,999,965, and 5,504,894 all describe resource demand distribution schemes that allocate client request among various resources where the client requests are each treated as independent events.

U.S. Pat. No. 6,173,322 is a good example of this approach and describes a system comprised of three host servers each having different request handling capabilities. For illustrative purposes, suppose that hosts H1, H2, and H3 have capabilities C1, C2, and C3 respectively with C3 being the most capable. Further suppose that there are three requests pending, R1, R2, and R3, needing capabilities C1, C2, and C3 respectively. If each request is considered independently and in the order the requests arrive, R1 might be assigned to H3 since this host will serve the request with the least delay. Next, R2 might be assigned to H2 for the same reason. R3 would then suffer if it were assigned to the only remaining host, H1, since H1 is under-powered to handle the request. Alternatively, R3 could wait for H3 to become available. The effect of these kinds of inefficiencies is cumulative; if the same three requests (or their respective equivalents) come in repeatedly and are serviced independently, there will be an ever-diminishing availability of resources until the system saturates and stops responding to new requests. Moreover, Internet demand is not well behaved. Service requests often come in bursts or may back up to form a large backlog for a variety of reasons. As a consequence, it is desirable for the resource allocation procedure to respond in a more sophisticated manner.

Another problem of the request distribution processes described in U.S. Pat. Nos. 6,070,191, 5,999, 965, and 5,504,894 is that these processes consider only parameters related to available resources and do not consider the attributes of the incoming client requests. U.S. Pat. No. 6,173,322 parses certain data contained in incoming clients requests, but only for the purpose of applying a static rule to distribute the requests to one of several server groups. Once this has been done, dynamic resource capability rules are applied to assign the request to a server within the group. These rules may operate in consideration of the static rules previously applied, but only after the static rules are first applied.

While existing schemes for distributing client requests among multiple servers have begun to address some of the problems that arise, it would be desirable to provide a system for distributing client requests across multiple servers that was more efficient and robust. Specifically, it would be advantageous to provide a system for distributing client requests across multiple servers that analyzed the attributes of client requests for expected demand patterns with which resource requirements may be associated, allowing for a comparison of the resource needs of incoming client requests with the resources available, and thus improving the capability of the resource allocation scheme to be more adaptive and dynamic from all operating aspects.

SUMMARY OF THE INVENTION

The present invention is a system for distributing incoming client requests across multiple servers in a networked client-server computer environment. The system collects information on both the attributes of the requests and the resource capability of the servers to dynamically allocate the requests in a set to the appropriate servers upon the completion of the time interval. Preferably, the system includes a request table to collect at least two requests incoming within a predetermined time interval. A request examiner routine analyzes each collected request with respect to at least one attribute. A system status monitor collects resource capability information of each server in a resource table at least once during said time interval. An optimization and allocation process distribute collected requests in the request table across the multiple servers upon completion of said time interval based on an optimization of potential pairings of the requests in the request table with the servers in the server table. The optimization and allocation process preferably analyzes metrics maintained in the request table and resource table as part of a relational database to allocate requests to servers based on a minimization of the metric distance between pairings of requests and servers. Preferably, the request table is part of a dynamic, relational database and a process of statistical inference for ascertaining expected demand patterns involving the attributes adds predictive information about client requests as part of the request examiner routine.

The present invention responds to the demanding circumstances described above by shifting from processing each request as an independent event to processing as a set of requests incoming within a predetermined time interval. The requests are processed as a set by collecting the requests incoming within the predetermined time interval, analyzing each of these requests with respect to at least one attribute, collecting at least once during the time interval information about each server's ability and availability, i.e., resource capability information, to handle requests, and distributing the set of requests across the multiple servers upon the completion of the time interval in response to the above actions, and then repeating these steps for each consecutive time interval. This invention has been denominated virtual extended technology (VXT) because it can intelligently run in the background within the confines of current day bandwidth and processing technology.

Resource allocation, the key to optimum throughput, is the real-time intelligent management of system resources. This invention utilizes several interactive decision processes that can consider all operating aspects of a system's resources, both static and dynamic, while balancing the continuously changing competition for these resources. One of the objectives of this invention is to provide a new algorithm for allocating Internet client requests in an intelligent manner to multiple servers to maximize the efficiency and fault tolerance of the resources. Costs of requests within a reasonable time interval are considered simultaneously to produce a solution that is globally effective (i.e., most effective for a site as a whole) at the possible expense of some individual (localized) requests. The objective is further achieved through analysis of attributes of requests as these attributes correlate to request demands on resources and of the just-in-time running status of those resources.

To return to the above example, a more effective solution would look at all three requests simultaneously, and assign R1 to H1, R2 to H2, and R3 to H3. The performance associated with request R1 will receive the nominal service (slightly less than that offered by the above solution) it needs, while R3 will also receive the appropriate level of service in a prompt manner, and the overall performance of the site will therefore be better. This latter solution is considered globally optimal because the number of requests managed per second is maximized and the collective resources are available for the next set of requests sooner.

Requests to a site can vary widely from web surfing, product search, price comparison, and checkout to multimedia access. However, the demand on resources by each kind of request is predictable. The distribution decision-making process of this invention accounts for attributes and behavior of incoming requests and the corresponding compatibility of system hardware and software. Incoming client requests are analyzed to determine their attributes and behavior so that a given request's expected demand on resources can be predicted and resource requirements can be assigned to the request.

One of the components of the invention will perform the extraction of the attributes from incoming requests. An analysis of the effectiveness of the characteristics above as well as the identification of other parameters that may be beneficial can be performed during the requirements analysis task. Extraction of the dynamic attributes will be performed in real-time by VXT's integral system request examiner or status monitor.

The invention learns how to characterize site-specific traffic in several ways. First, it expands or contracts the number of expected demand patterns based on the success of the request classification. In other words, if a live request does not sufficiently match an already existing pattern, a new pattern is created. Also, if the resource requirement parameters for the matching entry are not correct as measured by system experience, either the parameters themselves are adjusted, or a new pattern is created. Conversely, the number of patterns is constrained to minimize the computation required to classify live requests. The pattern set may be reorganized to eliminate unused, redundant, or ineffective entries. This self-organizing and reorganizing paradigm refines parameters by experience and remains vigilant to non-stationary statistical trends.

Similarly, the compatibility of the system hardware and software is also provided to the decision-making process. Some of these characteristics are static and known in advance, while others are dynamic and a function of the tasks currently executing. Preferably, a collection of resource capability information for each server includes metrics for CPU and memory availability and connectivity to a proxy server, to a main storage system, and to other content servers. This collection process can be push or pull from the server at certain times and any of several techniques can be implemented for minimal interruption of the main execution on the servers. For example, information can be pulled periodically by the main proxy server, or the main server can be pushed to accept such information from servers when any certain parameter exceeds a pre-determined threshold. This performance feedback allows for an informed decision on which request to send to which server.

Once this information is captured for a given interval of time, it must be reduced to a metric representation that can be manipulated to compute the best assignments of client requests to resources. The metrics associated with each request form a requirement data set whose elements represent the requirement level of each of the parameters used in the decision process. The metrics associated with the ability of a particular server to satisfy the request forms a capability data set with each element of this data set having a counterpart in the requirement data set. During operations, each request has its own requirement data set and each server or processing node has its own capability data set. The difference or metric distance between a requirement data set and a capability data set, calculated for any given pairing of client request and server, represents the mismatch (or cost) incurred by the corresponding assignment of the request to the server. If the data sets are identical, the cost is zero.

The assignment of multiple simultaneous requests can be done by one of several routines. The purpose of each routine, however, should be to select a server or processing resource for each client request so that the sum of all the costs, for the combination of resource and request pairings, is minimized. The solution can be found by using one of several algorithms.

Some algorithms find a perfect solution but require considerable processing, while others will find a nearly optimal solution quickly. Often, the nearly optimal solution is good enough to satisfy the presently existing circumstances.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagrammatic view of the present invention implemented across servers in a networked computing environment.

FIG. 2 is a diagrammatic view of the main interacting elements with a system with multiple servers for responding to client request, including a proxy server, content servers, switches, and storage system.

FIG. 3 illustrates five types of inter-processor connectivity for a system with multiple servers.

DESCRIPTION OF THE PREFERRED EMBODIMENT

FIG. 2 shows a typical configuration of a system (10) having multiple resources that may be allocated to respond to client requests received from the Internet. A proxy server(s) (12) receives the client request from the Internet and using the VXT (100), as will be described shortly, distributes those requests via network switches (14) to one of the multiple content/application servers (16) which preferably have access to a common DASD storage unit (18) on which information pertinent to the client requests may be stored. As will be discussed in connection with FIG. 3, it will be understood that the present invention is applicable to numerous configurations of server resources in a system (10). In one embodiment as described in the previously identified applications entitled “Scalable Internet Engine” and “Method and System For Providing Dynamic Host Service Management Across Disparate Accounts/Sites”, servers are dynamically allocated among multiple sites or accounts. In this embodiment, the present invention is applicable not only for allocating server resources among requests in a single account, but may also be extended to provide additional information for how to allocate servers among different accounts over time.

In the preferred embodiment of the VXT (100) as shown in FIG. 1, the invention comprises a request table (110) to collect at least two requests (102) incoming within a predetermined time interval, a request examiner process (120) to analyze each said collected request with respect to at least one attribute, a system status monitor (130) to collect resource capability information of each server (104), an optimization and allocation process (140) to distribute said collected requests in the request table (110) across the multiple servers upon completion of said time interval in response to said attributes and said resource capability information. Incoming client requests (102) are analyzed for their respective attributes by the request examiner process (120). This attribute information is then sent to the request table (110). Preferably, the system status monitor (130) collects resource capability information as part of a resource table (132).

In a preferred embodiment, the request table (110) and the resource table (132) are preferably implemented as part of a relational database. A process of rational statistical inference (150) analyzes each client request to assign a pattern classification so that its expected demand on resources can be predicted using the pattern classification in the adaptive request table (110).

One of the primary responsibilities of the request examiner (120) of the VXT (100) is to examine all incoming requests and to prioritize these requests based on criteria that can be described in general as (1) categorical criteria such as product searching, price, comparison, online shopping, web surfing, audio streaming, and video downloads, and (2) demographic criteria such as the origin of the request and possible user profile. Comparing these attributes with a dynamic, relational database that records past requests and their behavior along with a process of rational statistical inference (150) permits the VXT (100) to estimate each client request's (102) resource requirements in terms of CPU availability, memory availability, and bandwidth or connectivity of the servers (104).

The purpose of the database and process of statistical inference (150) is to facilitate the construction of an adaptive request table (110) containing several generic request types or pattern classifications that are most likely to be received by the proxy server (12). Each request type is assigned a set of at least five parameters or resource requirement metrics (114) that reflect different requirement aspects for the respective request. The values assigned to these five parameters form a requirements vector (116) that prescribes the generic request's expected resource requirements in terms of CPU time, memory, bandwidth or connectivity for storage, bandwidth or connectivity to the main proxy server, and bandwidth or connectivity to peer servers (i.e., connectivity between content servers). When a request from the Internet comes in, the request examiner (120) compares the request with the patterns (112) contained in the adaptive request table (110), finds the closest match, and creates a requirement vector (116) including the five corresponding resource parameters.

With reference to FIG. 3, a functional and cost effective system (10) should have at least two levels of networked servers. The lowest level consists of a collection of symmetric multiple processors (SMP) on the same bus. The bus should be a network having an aggregate bandwidth greater than 1 Gbps and very low latency. The next level is a collection of SMPs on one or more switches with less than 1 Gbps bandwidth and higher latency. The VXT (100) is designed to intelligently handle the added complexities of such an ASP system.

The VXT (100) ranks the available servers according to specific ranking criteria and servers' current running status in CPU availability, memory availability, storage connectivity, main proxy server connectivity, and pear server connectivity and generates a resource table (132) summarizing the resource capability metric (134) in a capability vector (136).

In a dynamic environment, each processor is capable of handling more than one task at a time, whether it is performing a price comparison or a search for a specific consumer item. The CPU availability parameter is defined in absolute terms as the unused portion of each processor's computing power measured in units of millions of instructions per second (MIPS).

The memory availability parameter is defined in absolute terms as the unused portion of each node's shared memory measured in units of megabytes divided by the number of processors in the node. This is because for SMP systems with several, processors (usually 4 to 8) in each node, the amount of memory available to one particular processor cannot be determined as the memory is shared among all processors in the same node.

Connectivity is a complex matter. In most systems, each processor has five different communication partners. Latency (determined by hardware) and available bandwidth (determined by current utilization) should be ascertained for each of these partners. Figure four identifies five types of inter-processor connectivity with a wide range of latency and bandwidth. Type I connectivity is between processors on the same node. Type II connectivity is between processors on different nodes, but on the same switch. Type III connectivity is between processors on different switches (for which a new parameter should be introduced to represent the number of hops to reach the partner processor). Type IV connectivity is between the processor and the proxy server. Type V connectivity is between the processor and the main storage system. Presently most ASP systems are not sophisticated enough to take advantage of the inter-processor connectivity information, i.e., Types II and III connectivity, so VXT (100) combines Types II and III connectivity into an aggregate connectivity. Three parameters are defined to represent available main proxy connectivity, central storage connectivity, and peer server connectivity. These parameters are bandwidths measured in units of Mbps recorded by the system status monitor.

Once all this information for incoming Internet requests and system resources is captured for a given time interval, it must be reduced to form a metric representation that can be manipulated to compute the best assignments of requests to resources. The metrics associated with each request forms a requirement vector (116) whose elements represent the requirement level of each of the parameters used in the decision process. The metrics associated with the ability of a particular server (104) to satisfy the request (102) is referred to as a capability vector (136). Each element of this vector (136) has a counterpart in the requirement vector (116). During operations, each request (102) has its own requirement vector (116), and each server or processing node (104) has its own capability vector (136). The vector space distance between the requirement vector (116) and capability vectors (136) for any given pairing of request (102) and server (104) represents the degree of mismatch (cost) incurred by the corresponding assignment of the request to that server. If the vectors are identical, the cost is zero.

The assignment of multiple simultaneous requests (102) can be done in several ways. As described above, the preferred approach creates a requirement vector (116) for each request (102) and capability vector (136) for each resource (104). The distance vector between each pair of request to resource then becomes an element in a cost matrix whereby the row index is a request identifier and the column index is the resource identifier. The cost matrix is usually sparse since some assignments may be ruled out for simple reasons. A decision-making algorithm then selects a resource for each request so that the sum of all the costs in the matrix is minimized for all combinations of requests and resources. There are several minimization techniques available, such as general neural network techniques, simulated annealing methods and generic assignment algorithm approaches

The preferred algorithm provides a fast quasi-optimal solution to the distribution problem based on standard methods. One example of such a standard method is a neural network paradigm as described in U.S. Pat. No. 5,548,683, the disclosure of which is hereby incorporated by reference. Other examples of generic algorithm approach would be a greedy search solution algorithm. A greedy algorithm can be applied when the optimization problem is to decide whether or not to include some element from a given set. A greedy algorithm begins with no elements and sequentially selects an element from the feasible set of remaining elements by myopic optimization. (The elements could have been sorted by some criterion, such as associated weights.) This results in an optimal solution to the problem if, and only if, there is an underlying matroid structure (for example, a spanning tree). Other types of generic assignment algorithms would include auction algorithms or Munres algorithms.

Although the preferred embodiment has been described herein, numerous changes and variations can be made and the scope of the present invention is intended to be defined by the claims.

Claims

1. A method for allocating a server selected from a plurality of servers to client requests originating over a predefined time interval at a plurality of user accounts, the method comprising:

collecting a plurality of client requests that arrive within the predefined time interval wherein at least two of said client requests are serviceable by the server and wherein a first of said at least two of said client requests originates at a first user account and a second of said at least two of said client requests originates at a second user account;

determining a first value of a cost metric for a first set of client request-server pairings wherein said first set includes at least one client request-server pair with said server being paired with either said first or said second of said at least two client requests;

determining a second value of a cost metric for a second set of client request-server pairings wherein said second set includes at least one client request-server pair with said server being paired with both said first and said second of said at least two client requests; and

at the end of said time interval distributing said client requests according to one of said first and said second set of client request-server pairings based on said first and second values of said cost metric.

2. The method of claim 1 wherein the step of determining the value of a cost metric for a set of client request-server pairings comprises the steps of:

at the commencement of said predefined time interval, initializing a cumulative value to zero;

for each client request-server pair in the set of client request-server pairings,

a) creating a requirement vector corresponding to said client request;

b) creating a capability vector corresponding to said server;

c) calculating an inner product of said requirement vector and said capability vector and adding said inner product to the cumulative value and repeating steps a), b) and c) for all client request-server pairs in the set of client request-server pairings whereupon said cumulative value represents the value of the cost metric.

3. The method of claim 1 wherein the step of distributing said client requests further comprises distributing said client requests according to said first set of client requests-server pairings if said first value of the cost metric is lower than the second value of the cost metric otherwise distributing said client requests according to said second set of client requests-server pairings.

4. The method of claim 1 wherein the step of determining a value of a cost metric for a set of client request-server pairings further comprises the steps of:

initializing a set of client request-server pairings at a commencement of the predefined time interval;

a) selecting a client request-server pair to satisfy a selection criteria;

b) creating a requirement vector corresponding to said client request;

c) creating a capability vector corresponding to said server;

d) calculating a distance between the requirement vector and the capability vector and adding said distance to the cumulative value when said distance exceeds a match threshold value and repeating steps a), b), c) and d); and

e) adding said client request-server pair to said set of client request-server pairings when said distance exceeds a match threshold, said cumulative value is less than a cost threshold and said client request has arrived within said predefined time interval.

5. The method of claim 4 wherein said selection criteria comprises matching a client request with a server to generate at least one client request-server pairing belonging to one of said first set and said second set.

6. A method for distributing client requests across a plurality of servers in a client-server networked system, the method comprising:

selecting a time window;

collecting client requests arriving within said time window wherein said client requests include at least a first plurality of said client requests that originate at a first user account and at least a second plurality of client requests that originate at a second user account;

determining a first cost metric corresponding to a first set of client request-server pairing wherein at least one server is paired with at least one of said first plurality of said client requests and at least one of said second plurality of client requests;

determining a second cost metric corresponding to a second set of client request-server pairings wherein said second set is characterized by first and second disjoint subsets with all pairings that include client requests originating at the first user account belonging to the first subset and all pairings that include client requests originating at the second user account belonging to the second subset; and

selecting one of said first set of client request-server pairs and said second set of client request-server pairs based on a differential between said first cost metric and said second cost metric.

7. A system for distributing load within a client-server network, comprising:

a plurality of interconnected servers wherein each server is associated with a capability vector having at least one element associated with a resource expected to be requested by at least one of a plurality of incoming client requests;

a dynamic capability vector determining module adapted to generate a dynamic capability vector for each server of said plurality of interconnected servers, said dynamic capability vector representing an update to said capability vector such that the at least one element of the capability vector corresponds to an unused portion of the resource associated with the at least one element and measured at the commencement of one of a sequence of predefined time intervals;

a requirement vector determining module configured to generate a requirement vector for each incoming client request during the one of a sequence of predefined time intervals; and

a load balancing module for selectively pairing said plurality of interconnected servers with one or more of said plurality of incoming client requests so as to minimize a cost metric computed during the one predefined time interval in said sequence of predefined time intervals wherein said cost metric is a function of vector distances between said dynamic capability vectors and said requirement vectors associated with said servers and said client request pairs in said server-client request pairing.

8. The system of claim 7 wherein said load balancing module further comprises a plurality of instances of load balancing modules resident on an appropriate plurality of servers disposed at intermediate nodes forming a connectivity hierarchy of layers throughout said client-server network such that said cost metric is computed and minimized for at least one layer of server nodes corresponding to the same connectivity hierarchy whereby each incoming client request is satisfied by a plurality of servers and transmission paths.

9. A method for creating a fast lookup table to determine server nodes within a cluster of nodes for servicing a plurality of client requests incoming within a predetermined time interval, the method comprising:

a) creating an adaptive request table populated with a set of patterns wherein each pattern is associated with a generic request type that is most likely to be received by said server nodes;

b) upon receiving a client request within said predetermined time interval, finding a match-pattern in said set of patterns that best matches the client request;

c) using the match-pattern to generate a requirements vector for said client request;

d) associating each server with a capability vector that is refreshed with resources available on said each server during said predefined time interval;

e) computing a score metricizing a vector distance between said requirements vector and said capability vector;

f) looking up a server node in the cluster of nodes to service the client request based upon the score, and

g) distributing the client request to at least one server node based upon said score.

10. A method for allocating hosting-service resources to clients in at least one shared server, said method comprising:

discovering utilization patterns of said clients;

monitoring said clients to discover said utilization patterns;

providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

modeling dimensions for client user measures and said utilization patterns; and

allocating said resources to said clients dependent on said utilization patterns.

11. The method according to claim 10, further including packing said clients using stochastic vectors.

12. The method according to claim 11, wherein said packing utilizes at least one of a Roof Avoidance process, a Minimized Variance process, a Maximized Minima process, and a Largest Combination process.

13. The method according to claim 10, wherein said hosting-service resources relate to at least one hosting service comprising one of collaborative hosting services, commerce hosting services, and e-business hosting services.

14. The method according to claim 10, wherein said allocating affects a Quality of Service (QoS) guarantee.

15. The method according to claim 10, wherein said utilization patterns are dependent upon access rates of one or more websites, said access rates have periodicity on multiple time scales.

16. The method according to claim 15, wherein two or more clients are selected from a plurality of clients on the basis of complementarity, wherein said hosting-service resources are allocated to said selected two or more clients as a combination.

17. The method according to claim 16, wherein said allocating comprises selecting said two or more clients to be allocated to a server, said two or more selected clients each having a peak load that is substantially disjoint in time in relation to a peak load of the remaining other selected clients.

18. The method according to claim 16, wherein said allocated hosting-service resources include resources allocated exclusively to each of said selected two or more clients and shared resources allocated to said combination for use by said selected two or more clients.

19. The method according to claim 17, wherein N clients are selected and allocated to a server, N being an integer greater than or equal to two, said server being partitioned into N virtual servers, each client being exclusively allocated a corresponding one of said N virtual servers, excess capacity of said server beyond the capacity required to provide said N virtual servers is shared by said N clients.

20. An apparatus for allocating hosting-service resources to clients in at least one shared server, said apparatus including:

means for discovering utilization patterns of said clients;

means for monitoring said clients to discover said utilization patterns;

means for providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

means for modeling dimensions for client user measures and said utilization patterns; and

means for allocating said resources to said clients dependent on said utilization patterns.

21. The apparatus according to claim 20, wherein said hosting-service resources relate to at least one hosting service comprising one of collaborative hosting services, commerce hosting services, and e-business hosting services.

22. The apparatus according to claim 20, wherein said allocating means affects a Quality of Service (QoS) guarantee.

23. The apparatus according to claim 20, wherein said utilization patterns are dependent upon access rates of one or more websites, said access rates have periodicity on multiple time scales.

24. The apparatus according to claim 23, wherein two or more clients are selected from a plurality of clients on the basis of complementarity, wherein said hosting-service resources are allocated to said selected two or more clients as a combination.

25. The apparatus according to claim 24, wherein said allocating means includes means for selecting said two or more clients to be allocated to a server, said two or more selected clients each having a peak load that is substantially disjoint in time in relation to a peak load of the remaining other selected clients.

26. The apparatus according to claim 25, wherein N clients are selected and allocated to a server, N being an integer greater than or equal to two, said server being partitioned into N virtual servers, each client being exclusively allocated a corresponding one of said N virtual servers, excess capacity of said server beyond the capacity required to provide said N virtual servers is shared by said N clients.

27. The apparatus according to claim 24, wherein said allocated hosting-service resources include resources allocated exclusively to each of said selected two or more clients and shared resources allocated to said combination for use by said selected two or more clients.

28. A computer program product having a computer readable medium having a computer program recorded therein for allocating hosting-service resources to clients in at least one shared server, said computer program product including:

computer program code means for discovering utilization patterns of said clients; and

computer program code means for monitoring said clients to discover said utilization patterns;

computer program code means for providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

computer program code means for modeling dimensions for client user measures and said utilization patterns; and

computer program code means for allocating said resources to said clients dependent on said utilization patterns.

29. A decision support system for allocating and planning resources in hosting computing services, said decision support system including:

means for modeling utilization of resources of one or more servers by clients in response to at least one of utilization patterns of said clients and specified rules regarding quality of service;

means for monitoring said clients to discover said utilization patterns;

means for providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

means for modeling dimensions for client user measures and said utilization patterns; and

means for determining a minimum number of servers for accommodating said clients to ensure a specified minimum quality of service.

30. The decision support system according to claim 29, wherein said determining means utilized stochastic vector packing.

31. The decision support system according to claim 29, wherein said system facilitates optimal management of resources in said hosting computing services.

32. The decision support system according to claim 29, wherein said hosting computing services include hosting computing resources, computing applications, computing-related services and network bandwidth.

33. The decision support system according to claim 29, including means for generating for a service provider a set of suggestions for optimal resource planning and allocation.

34. The decision support system according to claim 29, wherein said system provides an optimization service for use in a business model hosting optimization applications.

35. A decision support method for allocating and planning resources in hosting computing services, said method comprising:

modeling utilization of resources of one or more servers by clients in response to at least one of utilization patterns of said clients and specified rules regarding quality of service;

monitoring said clients to discover said utilization patterns;

providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

modeling dimensions for client user measures and utilization patterns; and

determining a minimum number of servers for accommodating said clients to ensure a specified minimum quality of service.

36. A computer program product having a computer readable medium having a computer program recorded therein for providing decision support to allocate and plan resources in hosting computing services, said computer program product including:

computer program code means for modeling utilization of resources of one or more servers by client in response to at least one of utilization patterns of said clients and specified rules regarding quality of service;

computer program code means for monitoring said clients to discover said utilization patterns;

computer program code means for providing bounds specifying minimum and maximum hosting-service resources for each of said clients;

computer program code means for modeling dimensions for client user measures and said utilization patterns; and

computer program code means for determining a minimum number of servers for accommodating said clients to ensure a specified minimum quality of service.

37. The computer program product according to claim 36, wherein said computer program code means for determining utilizes stochastic vector packing.

38. The computer program product according to claim 36, wherein said computer program product facilitates optimal management of resources in said hosting computing services.

39. The computer program product according to claim 36, wherein said hosting computing services include hosting computing resources, computing applications, computing-related services, and network bandwidth.

40. A method of improving load balancing operations in a computing network using cost metrics, comprising steps of:

obtaining cost metrics representing a cost of generating document content;

receiving a request for particular document content;

determining a particular one of a plurality servers which most recently served the requested document content; and

routing the request to a selected one of the plurality of servers, further comprising the steps of:

determining which other one of the plurality of servers is (1) capable of serving the requested document content and (2) most laghtly loaded;

using the obtained cost metrics to compare a first cost of routing the request to the determined one to a second cost of routine the request to the particular one; and

selecting the determined one if the first cost is less than the second cost and selecting the particular one otherwise.

41. The method according to claim 40, wherein the first cost and the second cost include a current load on the determined one and the particular one, respectively.

42. The method according to claim 40, wherein:

the obtaining step further comprises the step of receiving meta-data which conveys the cost metrics for the document content; and

the using step further comprises the step of using the cost metrics from the received meta-data.

43. The method according to claim 42, wherein the received meta-data comprises a HyperText Transfer Protocol (“HTTP”) response header.

44. The method according to claim 42, wherein the received meta-data comprises a plurality of HyperText Transfer Protocol (“HTTP”) response headers, each of the headers conveying an element of the cost metric for a particular document content.

45. The method according to claim 42, wherein the syntax comprises a specially denoted comment.

46. The method according to claim 40, wherein the markup language is XML (“Extensible Markup Language”).

47. A system for improving load balancing operations in a computing network using cost metrics, comprising:

means for obtaining cost metrics representing a cost of generating document content;

means for receiving a request for particular document content;

means for responding to the request using cached content, if available; and

means for routing the request to a selected one of a plurality of server; when cached content is not available, further comprising:

means for determinig a particular one of the plurality of servers that most recently served the requested document content;

means for determinig which other one of the plurality of servers is (1) capable of serving the requested document content and (2) most lightly loaded;

means for using the obtained cost metrics to compare a first cost of routing the request to the determined one to a second cost of routing the request to the particular one and means for selecting the determined one if the first cost is less than the second cost and selecting the particular one otherwise.

48. A computer program product for improving load balancing operations in a computing network using cost metrics, the computer program product embodied on one or more computer-readable media and comprising:

computer-readable program code means for obtaining cost metrics representing a cost of generating document content;

computer-readable program code means for receiving a request for particular document content;

computer-readable program code means for responding to the request using cached content, if available; and

computer-readable program code means for routing the request to a selected one of a plurality of servers, when cached content is not available, further comprising:

computer-readable program code means for determining a particular one of the plurality of servers that most recently served the requested document content;

computer-readable program code means for determining which other one of the plurality of servers is (1) capable of serving the requested document content and (2) most lightly loaded;

computer-readable program code means for using the obtained cost metrics to compare a first cost of routing the request to the determined one to a second cost of routing the request to the particular one; and

computer-readable program code means for selecting the determined one if the first cost is less than the second cost and selecting the particular one otherwise.

49. A method of using cost metrics when load balancing incoming content requests in a networking environment, comprising steps of:

gathering cost metric information representing a cost of generating document content; and

creating meta-data to convey the cost metric information to a load balancer;

sending the created meta-data to the load balancer;

receiving the sent cost metric information at the load balancer;

upon receiving a request for the document content at the load balancer, using the received cost metric information to route the request to a server selected from a plurality of servers,

further compromising the steps of:

using the received cost metric information to determine a first cost of serving the requested document content from a particular one of the plurality of servers that most recently served the requested document content;

using the received cost metric information to determine a second cost of serving the requested document content from a different one of the plurality of servers, wherein the received cost metric information indicates that the different one of the plurality of servers which is capable of serving the request document content at least cost; and

selecting the particular one if the first cost is less than the second cost and selecting the different one otherwise.

50. The method according to claim 49, wherein the gathered cost metric information comprises at least one of: (1) processing time at one of the plurality of servers which is an origin server; (2) network costs from the origin server to one or more ones of the plurality of servers which are backend enterprise servers; (3) processing time at the backend enterprise servers; and (4) a cost of delivering the generated document content to a proxy or cache.